diff --git a/python/paddle/base/dygraph/tensor_patch_methods.py b/python/paddle/base/dygraph/tensor_patch_methods.py
index a612a7106d5f2b..d8c589435cc3b4 100644
--- a/python/paddle/base/dygraph/tensor_patch_methods.py
+++ b/python/paddle/base/dygraph/tensor_patch_methods.py
@@ -127,6 +127,7 @@ def _to_static_var(self, to_parameter=False, **kwargs):
             'strides',
             'offset',
             '__cuda_array_interface__',
+            '__dlpack__',
         ]
         param_keys = ['stop_gradient', 'trainable']
         if isinstance(self, EagerParamBase):
@@ -1330,6 +1331,52 @@ def __cuda_array_interface__(self):
             "version": 2,
         }
 
+    def __dlpack__(self, stream=None):
+        """
+        Creates a DLPack capsule of the current tensor to be exported to other libraries.
+
+        Args:
+            stream (int | None): An optional Python integer representing a pointer
+                                to a CUDA stream. Synchronizes the tensor with this
+                                stream before exporting.
+                                If None or -1, no synchronization is performed.
+                                If 0, the default stream is used.
+        """
+        if "gpu" not in str(self.place):
+            raise AttributeError(
+                "Can't get __dlpack__ on non-CUDA tensor. "
+                "Use tensor.cuda() to move the tensor to device memory."
+            )
+
+        if self.is_sparse():
+            raise AttributeError(
+                "Can't get __dlpack__ on sparse tensor. "
+                "Use Tensor.to_dense() to convert to a dense tensor first."
+            )
+
+        if not self.stop_gradient:
+            raise RuntimeError(
+                "Can't get __dlpack__ on Tensor that requires gradients. "
+                "If gradients aren't required, use tensor.detach() to get a tensor without gradient."
+            )
+
+        if stream is not None and not isinstance(stream, int):
+            raise TypeError("stream must be an integer or None")
+
+        if stream is not None and stream != -1:
+            if self.place.is_gpu_place():
+                if stream == 0:
+                    stream = paddle.device.cuda.default_stream()
+                else:
+                    stream = paddle.device.cuda.ExternalStream(stream)
+                current_stream = paddle.device.cuda.current_stream()
+                if stream != current_stream:
+                    event = paddle.device.cuda.Event()
+                    event.record(current_stream)
+                    stream.wait_event(event)
+
+        return paddle.utils.dlpack.to_dlpack(self)
+
     if not hasattr(core, "eager"):
         return
 
@@ -1374,6 +1421,7 @@ def __cuda_array_interface__(self):
         ("_use_gpudnn", _use_gpudnn),
         ("_md5sum", _md5sum),
         ("__cuda_array_interface__", __cuda_array_interface__),
+        ("__dlpack__", __dlpack__),
     ):
         setattr(core.eager.Tensor, method_name, method)
 
diff --git a/test/dygraph_to_static/test_tensor_attr_consistency.py b/test/dygraph_to_static/test_tensor_attr_consistency.py
index d39494a84a559b..54952e58942067 100644
--- a/test/dygraph_to_static/test_tensor_attr_consistency.py
+++ b/test/dygraph_to_static/test_tensor_attr_consistency.py
@@ -78,6 +78,7 @@
         'value',
         'zero_',
         "__cuda_array_interface__",
+        "__dlpack__",
     ]
 )
 STATIC_ONLY_TENSOR_ATTRS_ALLOW_LIST = OrderedSet(
diff --git a/test/legacy_test/test_dlpack.py b/test/legacy_test/test_dlpack.py
index 064318f19e886b..6924e72250a965 100644
--- a/test/legacy_test/test_dlpack.py
+++ b/test/legacy_test/test_dlpack.py
@@ -325,6 +325,15 @@ def test_to_dlpack_from_zero_size(self):
                     np.testing.assert_array_equal(x.numpy(), y1.numpy())
                     np.testing.assert_array_equal(x.numpy(), y2.numpy())
 
+    def test_dlpack_basic(self):
+        tensor = paddle.to_tensor([1.0, 2.0, 3.0])
+        dlpack_capsule = tensor.__dlpack__()
+        self.assertIsNotNone(dlpack_capsule)
+
+        converted_tensor = paddle.from_dlpack(dlpack_capsule)
+        self.assertTrue(paddle.equal_all(tensor, converted_tensor))
+        self.assertEqual(tensor.data_ptr(), converted_tensor.data_ptr())
+
 
 class TestRaiseError(unittest.TestCase):
     def test_to_dlpack_raise_type_error(self):
diff --git a/third_party/pybind b/third_party/pybind
index a2e59f0e706540..3e9dfa28669416 160000
--- a/third_party/pybind
+++ b/third_party/pybind
@@ -1 +1 @@
-Subproject commit a2e59f0e7065404b44dfe92a28aca47ba1378dc4
+Subproject commit 3e9dfa2866941655c56877882565e7577de6fc7b
diff --git a/tools/enforce/grep_invalid_enforce.sh b/tools/enforce/grep_invalid_enforce.sh
index 04243bfb9afaf0..9e653c6b90fcc4 100644
--- a/tools/enforce/grep_invalid_enforce.sh
+++ b/tools/enforce/grep_invalid_enforce.sh
@@ -17,14 +17,14 @@
 # This script is used to grep invalid PADDLE checks by directory or file in the paddle/fluid/,
 #   the result show all invalid PADDLE checks in specified directory or file.
 
-# Usage: 
+# Usage:
 #   - bash grep_invalid_enforce.sh [target directory or file] (run in tools directory)
 #       - The default check path is paddle/fluid/operators
 
 # Result Examples:
 # 1. grep invalid PADDLE checks in directory
 
-#     - Command: /work/paddle/tools {develop} bash grep_invalid_enforce.sh ../paddle/fluid/imperative 
+#     - Command: /work/paddle/tools {develop} bash grep_invalid_enforce.sh ../paddle/fluid/imperative
 #     - Results:
 #         - paddle/fluid/imperative/gradient_accumulator.cc
 #         PADDLE_ENFORCE_EQ(dst_tensor->numel() == numel, true,
@@ -60,7 +60,7 @@
 #                     "Place cannot be CUDAPlace when use_double_buffer is False");
 #         PADDLE_ENFORCE_NOT_NULL(exceptions_[i]);
 #         PADDLE_ENFORCE_EQ(status, Status::kException);
-#         PADDLE_ENFORCE_EQ(status, Status::kSuccess);    
+#         PADDLE_ENFORCE_EQ(status, Status::kSuccess);
 
 . ./count_enforce_by_file.sh --source-only